AI Models Predict Arsenic in Bihar Groundwater With Bio-Inspired Precision

In the floodplain state of Bihar, India, the water that sustains tens of millions of people carries an invisible threat. Arsenic, leached naturally from sediments into shallow aquifers, has contaminated groundwater across much of the state, and chronic exposure is linked to cancers, cardiovascular disease, and neurological impairment. Because groundwater is the primary source of daily drinking water in Bihar, knowing where dangerous concentrations occur before people drink them is a public health imperative. A new study from researchers at the National Institute of Technology Patna, published in Earth Science Informatics, tackles this challenge with a pair of hybrid machine learning frameworks that couple the popular gradient boosting algorithm XGBoost with nature-inspired optimization routines, achieving strikingly high predictive accuracy on the data available.

The research, led by Abhishek Kumar Mishra with Furquan Ahmad and Nityanand Singh Maurya, addresses a persistent problem in arsenic risk assessment: direct chemical testing of every well is slow, expensive, and logistically difficult across a densely populated state with millions of private and public extraction points. Predictive models offer a way to screen locations cheaply, flagging wells and districts where follow-up sampling is most urgent. The team built two models, named XGB-MPA and XGB-OOA, both of which use Extreme Gradient Boosting as the underlying learner but differ in how the algorithm’s internal settings are tuned. XGB-MPA employs the Marine Predators Algorithm, while XGB-OOA uses the Osprey Optimization Algorithm, both relatively recent metaheuristics that mimic biological foraging behavior to search vast parameter spaces efficiently.

XGBoost itself is an ensemble technique that builds a sequence of decision trees, each new tree trained to correct the residual errors of its predecessors. Its predictive power depends heavily on hyperparameters such as tree depth, learning rate, and regularization strengths, and poor choices can leave substantial performance on the table. That is where the bio-inspired optimizers come in. The Marine Predators Algorithm, introduced in 2020, models the interaction between predators and prey in ocean ecosystems, alternating between Brownian and Levy flight movement patterns to balance exploration of the search space with exploitation of promising regions. The Osprey Optimization Algorithm, published in 2023, simulates how ospreys locate and capture fish, similarly guiding candidate solutions toward optimal configurations. By letting these algorithms automatically tune XGBoost’s settings, the researchers removed much of the guesswork that typically burdens machine learning applications in hydrology.

The models were trained on four major hydrogeochemical and depth-related variables: pH, electrical conductivity, total dissolved solids, and well depth. These parameters are routinely measured and are physically meaningful indicators of the geochemical conditions that govern arsenic mobilization in aquifer sediments. Arsenic release in the alluvial aquifers of the Ganges basin is closely tied to redox conditions, sediment chemistry, and groundwater flow paths, all of which leave fingerprints in the basic water chemistry that the models can learn to interpret. The target of prediction was the arsenic concentration itself, allowing the frameworks to estimate contamination levels rather than merely classify wells as safe or unsafe.

Performance evaluation was unusually thorough for a study of this kind. The researchers relied on a battery of statistical metrics, including the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), the variance account for factor (VAF), and the ratio of root mean square error to the standard deviation of measured data (RSR). Beyond these, they subjected the models to k-fold cross-validation to test stability across different data partitions, applied the Wilcoxon signed-rank test to confirm that differences between the two models were statistically significant rather than artifacts of random variation, and used Anderson-Darling probability analysis to examine the distribution of prediction errors. This layered validation strategy is designed to guard against the most common failure mode of machine learning studies: models that look impressive on paper but collapse when confronted with new data.

The results favored the Marine Predators Algorithm variant. XGB-MPA achieved an R² of 0.9752 on training data and 0.9705 on testing data, with normalized RMSE values of 0.0370 and 0.0325 respectively, edging out XGB-OOA in predictive accuracy and generalization capacity. An R² approaching 0.97 means the model explains nearly all of the variance in observed arsenic concentrations within the dataset, a level of agreement that, if it holds up under independent testing, would represent a powerful screening tool. Complementary indices reinforced the picture: the a20 index, which measures the fraction of predictions falling within a factor of deviation of observed values, reached 84.29 percent for training and 76.67 percent for testing, while the index of agreement (IOA) exceeded 0.98 and the index of scatter (IOS) stayed below 0.19. Together these figures indicate that XGB-MPA’s predictions track measured concentrations closely and consistently across both phases of model development.

The stakes in Bihar could hardly be higher. State economic survey data cited in the study report massive groundwater contamination affecting 31 of the state’s 38 districts, and the World Health Organization’s drinking water guidelines identify arsenic as one of the most significant inorganic contaminants globally. Chronic exposure through drinking water is associated with skin lesions, cancers of the skin, bladder, and lung, heart disease, and cognitive and neurological deficits. Because arsenic is tasteless, colorless, and odorless, residents typically have no way of knowing their exposure level without chemical testing. A reliable predictive model, even one used only as a first-pass screening instrument, could direct scarce testing resources to the places where they will do the most good, potentially shortening the lag between contamination and intervention.

The authors are careful, and appropriately so, about the limits of what their frameworks can claim. They position the models as a data-driven preliminary screening aid for identifying locations that require further groundwater-quality investigation, not as a replacement for direct measurement. Operational water-management decisions and health-related interventions, they stress, should incorporate actual water-quality measurements and a broader set of hydrogeochemical indicators, including iron and manganese, which interact with arsenic in aquifer systems and were not part of the model’s input set. They also note that broader transferability of the framework requires validation using independent datasets that incorporate additional hydrogeochemical variables, a candid acknowledgment that the impressive statistics reported here describe performance within the investigated dataset rather than guaranteed performance across the full hydrological diversity of the Ganges plain.

That caveat matters because arsenic prediction is a notoriously difficult modeling problem. Concentrations can vary dramatically between wells separated by only a few hundred meters, reflecting the patchwork of buried organic matter, iron oxide minerals, and sedimentary history that controls where arsenic is released into solution. Models trained in one geomorphic setting often falter in another. The Bihar study’s use of only four input variables is both a strength and a limitation: it makes the screening approach practical, since pH, conductivity, total dissolved solids, and depth are cheap and widely measured, but it also means the models capture only part of the geochemical complexity that determines arsenic fate. Future iterations that fold in iron, manganese, redox indicators, and perhaps remote sensing or geological layers could sharpen predictions further and extend their reach.

Still, the study adds to a growing body of evidence that hybrid machine learning, in which gradient boosting ensembles are tuned by metaheuristic optimizers, can extract reliable predictive signal from modest environmental datasets. The head-to-head comparison of two recent optimization algorithms, validated with nonparametric statistical testing, offers a template for how such models should be evaluated before being proposed for real-world use. For the millions of Bihar residents who draw their drinking water from arsenic-prone aquifers, the practical promise is straightforward: a fast, data-driven way to identify which wells deserve urgent testing and which communities face the greatest hidden risk. If independent validation confirms what the training statistics suggest, frameworks like XGB-MPA could become a standard first line of defense in the ongoing effort to keep arsenic out of the region’s drinking water.

Subject of Research: Hybrid machine learning prediction of groundwater arsenic contamination in Bihar, India

Article Title: Optimization-driven hybrid XGBoost frameworks for predicting groundwater arsenic contamination in Bihar, India

Article References: Mishra, A. K., Ahmad, F., & Maurya, N. S. (2026). Optimization-driven hybrid XGBoost frameworks for predicting groundwater arsenic contamination in Bihar, India. Earth Science Informatics, 19(11), Article 207. https://doi.org/10.1007/s12145-026-02262-z

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02262-z

Keywords: groundwater, arsenic contamination, XGBoost, machine learning, Marine Predators Algorithm, Osprey Optimization Algorithm, Bihar, hydrogeochemistry, drinking water safety, bio-inspired optimization, water quality prediction, India